Hook
Last week, a viral analysis of Zhipu AI—the Chinese startup behind the GLM series of large language models—painted a picture that struck a chord with anyone who has watched the blockchain space long enough. Zhipu AI runs a free model, burns through cash at an alarming rate, and yet enjoys a valuation that briefly surpassed that of Meituan, a profitable food-delivery giant. Sound familiar? Replace "free model" with "zero-gas transactions" and "billion-dollar valuation" with "fully diluted valuation of a Layer 1," and you have the story of half the projects in crypto today. As someone who spent 2017 manually auditing whitepapers during the ICO boom, I’ve learned that technical brilliance without a sustainable economic base is a house built on sand. Zhipu AI’s story is not just about AI—it’s a mirror held up to our own industry’s deepest flaws.
Context
For readers who haven’t followed the Chinese AI landscape closely, Zhipu AI is the crown jewel of the Tsinghua University lineage, developing the GLM architecture from scratch rather than fine-tuning someone else’s model. The company made headlines by offering its most capable models for free—no API costs, no tiered pricing—while simultaneously reporting massive losses from training and inference compute. The analysis I reviewed concluded that Zhipu AI’s high valuation is a bet on future monopoly power, not present profitability. The parallels with blockchain are irresistible: from Ethereum’s early “world computer” narrative to Solana’s “ecosystem first, monetization later” strategy, we’ve celebrated projects that spend heavily to capture market share, trusting that scale will eventually unlock revenue. But Zhipu AI’s free model comes with variable costs that scale linearly with usage—just like a blockchain’s gas fees, node rewards, or sequencer expenses. The critical question: is there a path to positive unit economics, or are we subsidizing a mirage?
Core
To understand the core tension, we need to dive into the technical and economic geometry of both AI and blockchain. Based on my audit experience with DeFi protocols during the 2020 Trust Repair Workshops, I’ve seen firsthand how projects hide their cost structures behind vague tokenomics. Zhipu AI’s situation is more transparent but equally painful: each inference request consumes GPU compute, which costs real money. Even with optimized models, the marginal cost of serving a free user on a large LLM is measurable in cents—and when those users multiply by millions, the burn accelerates exponentially.
In blockchain, the analogous cost is validation. A proof-of-work chain consumes electricity as a function of hash rate; a proof-of-stake chain may have lower energy costs but still incurs substantial opportunity costs for stakers. When projects like Arbitrum or Optimism offered free transaction subsidies during their early days, they were effectively paying users’ gas fees from a treasury. That’s no different from Zhipu AI paying for inference compute. The difference is that blockchain projects often mask this via token inflation: they print tokens to pay validators or subsidize users, deferring the cost to future holders. Zhipu AI, being a private company, has to burn real cash from venture capital. Both models assume that adoption will eventually create a revenue stream large enough to cover these costs—but the evidence suggests otherwise.
Let’s turn to the numbers. Zhipu AI’s free model is not a stripped-down version; it’s the same architecture that scores well on Chinese benchmarks. The company reportedly spends tens of millions of dollars annually on compute. If they were to charge even a fraction of OpenAI’s API rates, they could break even—but they don’t. Why? Because the market is brutally competitive. Baidu, Alibaba, and Tencent also offer free AI models, and startups like Moonshot AI (maker of Kimi) are winning on user experience. Zhipu AI is trapped in a race to the bottom, where the only differentiator is—you guessed it—being free for longer. In blockchain, we see the same dynamic: Layer 2s race to zero gas fees, DEXs compete on zero-slippage promises, and NFT marketplaces waive creator royalties. The subsidized user is not a loyal user; they are a mercenary who will leave as soon as the subsidy ends.
I recall a project I audited in 2021: a gaming NFT platform that let players “earn” tokens by completing tasks, with no real utility beyond speculation. The team argued that “adoption first, monetization later” was the path to success. Six months later, when the treasury dried up, the token collapsed, and the community evaporated. Zhipu AI faces the same risk, but with a safety net of geopolitical importance—Beijing sees it as a strategic asset. Most blockchain projects lack that cushion. The technical insight here is that both AI models and blockchain protocols are infrastructure with high fixed costs (training/development) and low marginal costs (inference/transaction), but the marginal cost is never zero. Free forever is a lie.
Contrarian
Now, let me flip the script. Perhaps Zhipu AI’s strategy isn’t as foolish as it seems. By offering free models, they are building a developer ecosystem that will be sticky once they introduce paid tiers—similar to how Hugging Face monetizes through enterprise compute after giving away model weights. In blockchain, we see this with projects like Ethereum: the base layer is permissionless (free to build on), but scaling solutions like rollups charge fees. The contrarian view is that Zhipu AI is playing the long game, acquiring users who will later pay for premium features like higher rate limits, dedicated compute, or model fine-tuning. The key to this strategy is differentiation: if your free model is indistinguishable from competitors, you have no pricing power. Zhipu AI’s GLM architecture has unique advantages in Chinese-language tasks and regulatory compliance, which could give it an edge in government and enterprise contracts. Similarly, blockchain projects like Solana differentiate on throughput, and those that survive the subsidy phase often become indispensable for specific use cases (e.g., NFT minting on Solana). The blind spot in the Zhipu AI analysis is assuming that free means undifferentiated. It doesn’t—it means the product is a hook, not the full offering.
But here’s where blockchain must learn from Zhipu AI’s red flags. The company’s high valuation is based on a future monopoly that may never materialize, especially with state-backed competitors and new startups leapfrogging in product experience. In crypto, we see the same: Binance Smart Chain (now BNB Chain) built a massive ecosystem on cheap fees, but when Ethereum’s Layer 2s caught up in speed and cost, BNB Chain’s dominance began to erode. The contrarian angle is that free models can work if the project builds a network effect that outlasts the subsidy—but the timeline is measured in months, not decades. Zhipu AI needs to start charging within 12–18 months or risk running out of runway. Most blockchain projects with similar strategies have already failed; the few that survived (e.g., Uniswap with its fee switch) did so because they had a clear governance path to monetization.
Takeaway
What does Zhipu AI’s story mean for us in blockchain? It reinforces a principle I’ve believed since the 2017 ICO audit: technical excellence without sustainable economics is not a revolution—it’s an expensive hobby. The next time you see a blockchain project boasting about zero fees or free airdrops, ask yourself: where is the revenue coming from? If the answer is “adoption first, monetization later,” you are betting on the same mirage that inflates AI company valuations. The market will eventually demand that every protocol, like every AI model, proves its unit economics. Building bridges where code ends and trust begins requires more than just smart contracts—it requires a business model that respects the human cost of infrastructure. Auditing ethics before auditing assets means asking not just “is the code secure?” but “is the economics honest?”. Humanity is the ultimate protocol, and communities deserve projects that don’t burn through their trust alongside their treasury. Let Zhipu AI serve as a cautionary tale: free is fine for a demo, but for a future, you need a foundation that values sustainability over hype.
